Quantitative Argumentation Debate Semantics

• QuAD semantics is a quantitative framework that assigns each argument an initial strength and computes recursive acceptability scores by aggregating attack and support relations.
• It employs modular aggregation and influence functions, including product-based and piecewise-linear mechanisms, to balance conflicting evidence.
• QuAD supports efficient evaluation in acyclic frameworks and serves as a formal benchmark in debate analysis and structured non-linear reasoning.

Searching arXiv for QuAD semantics and related QBAF papers to ground the article. Quantitative Argumentation Debate (QuAD) semantics is a gradual semantics for Quantitative Bipolar Argumentation Frameworks (QBAFs) in which each argument has an initial strength and a recursively computed acceptability score determined by the combined effects of attack and support relations. In the supplied literature, QuAD appears as part of the modular-semantics landscape for bipolar argumentation, where an aggregation function combines incoming support and attack and an influence function maps the resulting signal together with the initial strength to an updated value (Alfano et al., 4 May 2026). The same literature also shows that terminology is not fully uniform: one source identifies DF-QuAD, often called QuAD, with a product-based aggregator and piecewise-linear influence function (Alfano et al., 4 May 2026), whereas another source presents a “QuAD (Quadratic-Energy)” gradual semantics in a broader modular family centered on a quadratic-energy influence rule (Alfano et al., 3 Mar 2026). This terminological variation is central to the interpretation of QuAD in current research.

1. Formal setting in quantitative bipolar argumentation

A Quantitative Bipolar Argumentation Framework is defined as a tuple

$\Delta = (A,\,R,\,S,\,\tau)$

where $A$ is a finite set of arguments, $R\subseteq A\times A$ is the attack relation, $S\subseteq A\times A$ is the support relation, and $\tau\colon A\to[0,1]$ assigns each argument an initial strength (Alfano et al., 4 May 2026). A related formulation used in a fuzzy argumentative knowledge base extends this structure to

$F=\<A,R,S,\tau,\varphi\>$

where $\varphi\colon R\cup S\to[0,1]$ assigns each attack or support link a fuzzy weight (Alfano et al., 3 Mar 2026).

Within this setting, a gradual semantics is an update function computed by iterated updates,

$$s^{(0)}_a=\tau_a,\qquad s_a^{(t+1)}=\iota\bigl(\alpha_a(s^{(t)}),\,\tau_a\bigr),$$

where $\alpha_a$ aggregates the impact of the attackers and supporters of $a$, and $A$0 determines how this impact modifies the initial strength (Alfano et al., 4 May 2026). The same modular decomposition is stated in the fuzzy-DL framework: aggregation computes the net incoming support minus attack, and influence transforms the pair $A$1 into the updated strength $A$2 (Alfano et al., 3 Mar 2026).

This modular perspective situates QuAD among a broader family of QBAF semantics. One source distinguishes DF-QuAD from sum-based semantics such as MLP, REB, and QEN: DF-QuAD uses a product-based aggregation mechanism, whereas MLP, REB, and QEN use $A$3 with different nonlinear influence functions (Alfano et al., 4 May 2026). This suggests that QuAD is best understood not merely as a single formula, but as a design point in a general architecture for propagating quantitative support and attack.

2. Core QuAD equations

In the formulation explicitly presented as QuAD semantics for debate graphs, a debate is modeled as a bipolar argumentation graph

$A$4

with base weights $A$5 for all $A$6 (Sanayei et al., 19 Sep 2025). For each argument $A$7, the sets of attackers and supporters are

$A$8

The semantics computes an acceptability score

$A$9

through intermediate attack and support valuations (Sanayei et al., 19 Sep 2025): $R\subseteq A\times A$0

The final acceptability is then given by the piecewise rule

$R\subseteq A\times A$1

In the experiments reported in that source, all base weights are initialized uniformly to $R\subseteq A\times A$2 (Sanayei et al., 19 Sep 2025).

A second presentation, given in a comparative discussion of modular semantics, characterizes DF-QuAD by the product-based aggregator

$R\subseteq A\times A$3

together with a piecewise-linear $R\subseteq A\times A$4 (Alfano et al., 4 May 2026). The relation between this representation and the previous one is that both are product-based formulations balancing attacks and supports recursively. A plausible implication is that QuAD’s characteristic behavior arises from multiplicative attenuation and amplification rather than from additive net-balance accumulation.

The supplied material also includes an alternative formula described as an equivalent presentation: $R\subseteq A\times A$5 with all $R\subseteq A\times A$6 and $R\subseteq A\times A$7 in that work (Sanayei et al., 19 Sep 2025). Because the source explicitly states that this is an equivalent presentation “in many presentations,” it is appropriately read as a common reformulation rather than as a replacement of the piecewise definition.

3. Iterative and topological computation

QuAD is computed recursively, either by a single topological sweep in acyclic graphs or by iterative fixed-point computation. In the debate-ranking formulation, because the graphs in both NoDE datasets are acyclic, QuAD converges to a unique fixed point by a single topological pass (Sanayei et al., 19 Sep 2025). The concrete procedure is: topologically sort the DAG, initialize $R\subseteq A\times A$8, compute $R\subseteq A\times A$9 and $S\subseteq A\times A$0 for each argument in that order, and then set $S\subseteq A\times A$1 by the piecewise equation (Sanayei et al., 19 Sep 2025). Since each argument’s new score depends only on its predecessors in the topological order, a single pass suffices (Sanayei et al., 19 Sep 2025).

The broader gradual-semantics framework expresses the same pattern through iteration: $S\subseteq A\times A$2 with convergence defined by the existence of the limit of the strength vector (Alfano et al., 4 May 2026, Alfano et al., 3 Mar 2026). A practical stopping criterion given in the fuzzy-DL account is

$S\subseteq A\times A$3

for some small $S\subseteq A\times A$4 (Alfano et al., 3 Mar 2026).

For acyclic QBAFs, the complexity reported in the supplied sources is linear in the graph size. One source states that in any finite acyclic QBAF one can topologically order the arguments so that each $S\subseteq A\times A$5 depends only on previously computed strengths, and hence in one pass one reaches the fixed point, with convergence in $S\subseteq A\times A$6 time (Alfano et al., 4 May 2026). Another source gives the same acyclic complexity bound for modular gradual semantics and adds that each general $S\subseteq A\times A$7-iteration costs $S\subseteq A\times A$8 (Alfano et al., 3 Mar 2026).

These facts place QuAD within the standard algorithmic profile of QBAF semantics: acyclic frameworks admit direct evaluation, while cyclic frameworks require fixed-point analysis or stronger assumptions on the update operator.

4. Worked examples and local behavior

The supplied literature gives compact examples that illustrate how QuAD propagates attack and support. In a three-argument toy debate with $S\subseteq A\times A$9 as support and $\tau\colon A\to[0,1]$0 as attack, with all $\tau\colon A\to[0,1]$1, the leaf arguments $\tau\colon A\to[0,1]$2 and $\tau\colon A\to[0,1]$3 remain at $\tau\colon A\to[0,1]$4 because they have neither attackers nor supporters (Sanayei et al., 19 Sep 2025). For $\tau\colon A\to[0,1]$5,

$\tau\colon A\to[0,1]$6

Since both attackers and supporters are present,

$\tau\colon A\to[0,1]$7

Thus the acceptability degrees after one iteration are

$\tau\colon A\to[0,1]$8

(Sanayei et al., 19 Sep 2025).

A second example, presented in the fuzzy-DL setting for the quadratic-energy update rule, considers three arguments $\tau\colon A\to[0,1]$9 where $F=\<A,R,S,\tau,\varphi\>$0 attacks $F=\<A,R,S,\tau,\varphi\>$1, $F=\<A,R,S,\tau,\varphi\>$2 supports $F=\<A,R,S,\tau,\varphi\>$3, $F=\<A,R,S,\tau,\varphi\>$4, $F=\<A,R,S,\tau,\varphi\>$5, $F=\<A,R,S,\tau,\varphi\>$6, and all attack/support weights are $F=\<A,R,S,\tau,\varphi\>$7 (Alfano et al., 3 Mar 2026). The source computes

$F=\<A,R,S,\tau,\varphi\>$8

then

$F=\<A,R,S,\tau,\varphi\>$9

and because $\varphi\colon R\cup S\to[0,1]$0,

$\varphi\colon R\cup S\to[0,1]$1

The second iteration is unchanged, so the fixed point is

$\varphi\colon R\cup S\to[0,1]$2

(Alfano et al., 3 Mar 2026).

Taken together, these examples clarify two recurring themes. First, QuAD-style semantics is sensitive to the pattern of incoming relations rather than only to counts of supporters and attackers. Second, when the product-based combination is used, equal-strength support and attack can neutralize one another in a balanced way, as in the three-argument toy debate (Sanayei et al., 19 Sep 2025). This suggests that QuAD is particularly suited to settings where recursive relational structure, rather than isolated argument quality, is the primary object of analysis.

5. Theoretical properties, convergence, and common points of confusion

The supplied sources attribute standard fixed-point properties to modular QBAF semantics. Because $\varphi\colon R\cup S\to[0,1]$3 is continuous, Brouwer’s fixed-point theorem guarantees the existence of at least one fixed point $\varphi\colon R\cup S\to[0,1]$4 (Alfano et al., 3 Mar 2026). In acyclic frameworks, a unique solution is obtained by topological evaluation in linear time (Alfano et al., 3 Mar 2026). In cyclic frameworks, uniqueness and convergence require additional assumptions, such as Lipschitz continuity with constant $\varphi\colon R\cup S\to[0,1]$5 for the aggregation–influence combination (Alfano et al., 3 Mar 2026).

A central point of confusion in the supplied material concerns the scope of QuAD convergence guarantees. The debate-evaluation study states that only acyclic graphs were considered and that “QuAD convergence is only guaranteed in that case” (Sanayei et al., 19 Sep 2025). By contrast, the dReLU paper analyzes its own convergence behavior much more broadly and uses this as part of its comparison with QuAD and related semantics (Alfano et al., 4 May 2026). The two statements are not contradictory: one limits the experimental benchmark to acyclic graphs, while the other positions QuAD as a baseline whose behavior can be problematic outside that setting.

Another important confusion concerns naming. One source explicitly distinguishes “DF-QuAD (QuAD for short)” from sum-based semantics such as MLP, REB, and QEN (Alfano et al., 4 May 2026). Another source refers to a “QuAD (Quadratic-Energy) gradual semantics” and then defines the update via the quadratic-energy function

$\varphi\colon R\cup S\to[0,1]$6

(Alfano et al., 3 Mar 2026). For an encyclopedia treatment, the safest conclusion is that “QuAD” is used non-uniformly across the supplied literature. The product-based, DF-QuAD characterization is the most explicit account of QuAD as a named semantics (Alfano et al., 4 May 2026), while the quadratic-energy formulation belongs to the same modular gradual-semantics ecosystem and is labeled differently in another source (Alfano et al., 3 Mar 2026).

6. Comparative position and use in debate analysis

The dReLU comparison paper places QuAD within a broader landscape of QBAF semantics and states that QuAD tends to treat equal-strength supporters and attackers in a “winner-take-all” way, sometimes yielding counter-intuitive results when many weak arguments accumulate (Alfano et al., 4 May 2026). The same source contrasts this with dReLU semantics, which rectifies positive and negative evidence separately and clamps the final update into $\varphi\colon R\cup S\to[0,1]$7 before renormalizing to $\varphi\colon R\cup S\to[0,1]$8 (Alfano et al., 4 May 2026). Examples in that paper are said to show that dReLU restores intuitive behavior in cases where QuAD or QEN give extreme or flat outcomes, such as when a single attack of strength $\varphi\colon R\cup S\to[0,1]$9 no longer freezes the attacked node at $$s^{(0)}_a=\tau_a,\qquad s_a^{(t+1)}=\iota\bigl(\alpha_a(s^{(t)}),\,\tau_a\bigr),$$0, or when a huge but nearly balanced crowd of supporters versus attackers yields a near-$$s^{(0)}_a=\tau_a,\qquad s_a^{(t+1)}=\iota\bigl(\alpha_a(s^{(t)}),\,\tau_a\bigr),$$1 result rather than a runaway toward $$s^{(0)}_a=\tau_a,\qquad s_a^{(t+1)}=\iota\bigl(\alpha_a(s^{(t)}),\,\tau_a\bigr),$$2 or $$s^{(0)}_a=\tau_a,\qquad s_a^{(t+1)}=\iota\bigl(\alpha_a(s^{(t)}),\,\tau_a\bigr),$$3 (Alfano et al., 4 May 2026).

In a separate line of work, QuAD is used as a formal target for evaluating whether LLMs can approximate structured non-linear reasoning over debates (Sanayei et al., 19 Sep 2025). In that setup, arguments are presented in dialogue form with attack and support edges stripped, and models must rank them without access to the underlying graph (Sanayei et al., 19 Sep 2025). The evaluation compares LLM-predicted rankings against QuAD rankings using Spearman’s rank correlation and Kendall’s $$s^{(0)}_a=\tau_a,\qquad s_a^{(t+1)}=\iota\bigl(\alpha_a(s^{(t)}),\,\tau_a\bigr),$$4, and also measures graph reconstruction quality under Chain-of-Thought prompts through precision, recall, and F1 (Sanayei et al., 19 Sep 2025).

The reported findings show moderate alignment with QuAD rankings and a strong dependence on prompting strategy and input characteristics (Sanayei et al., 19 Sep 2025). On DebatePedia, averaged across four LLMs, Vanilla prompting yields approximately $$s^{(0)}_a=\tau_a,\qquad s_a^{(t+1)}=\iota\bigl(\alpha_a(s^{(t)}),\,\tau_a\bigr),$$5, while CoT few-shot yields approximately $$s^{(0)}_a=\tau_a,\qquad s_a^{(t+1)}=\iota\bigl(\alpha_a(s^{(t)}),\,\tau_a\bigr),$$6 (Sanayei et al., 19 Sep 2025). On 12AngryMen, Vanilla yields approximately $$s^{(0)}_a=\tau_a,\qquad s_a^{(t+1)}=\iota\bigl(\alpha_a(s^{(t)}),\,\tau_a\bigr),$$7, while CoT few-shot yields approximately $$s^{(0)}_a=\tau_a,\qquad s_a^{(t+1)}=\iota\bigl(\alpha_a(s^{(t)}),\,\tau_a\bigr),$$8 (Sanayei et al., 19 Sep 2025). Performance degrades on longer inputs, randomizing argument order lowers performance by $$s^{(0)}_a=\tau_a,\qquad s_a^{(t+1)}=\iota\bigl(\alpha_a(s^{(t)}),\,\tau_a\bigr),$$9 and $\alpha_a$0, and CoT few-shot largely neutralizes argument-length and positional biases (Sanayei et al., 19 Sep 2025). A plausible implication is that QuAD provides a useful formal benchmark precisely because it exposes the gap between discourse-level fluency and graph-sensitive acceptability computation.

7. Extensions to fuzzy knowledge bases and description logics

QuAD-style quantitative reasoning is also embedded into a framework that integrates learning-based argument mining with ontology-based querying (Alfano et al., 3 Mar 2026). In that setting, a Fuzzy Argumentative ABox extends a standard DL ABox by allowing degree-annotated assertions of the forms $\alpha_a$1, $\alpha_a$2, and $\alpha_a$3, while all other concept or role assertions remain Boolean (Alfano et al., 3 Mar 2026). Proposition 3.5 of that source shows that any consistent FABox subsumes a QBAF

$\alpha_a$4

with arguments, attack links, support links, initial strengths, and link weights extracted directly from the ABox (Alfano et al., 3 Mar 2026).

Quantitative argumentation semantics are then applied to compute final argument strengths by propagating the effects of supports and attacks, and the results are embedded into a fuzzy description logic setting that enables expressive query answering through efficient rewriting techniques (Alfano et al., 3 Mar 2026). In this combined framework, consistency checking is in PTIME, and query answering remains in LOGSPACE with respect to the ABox size (Alfano et al., 3 Mar 2026).

This integration is significant because it treats QuAD-style reasoning not as an isolated scoring procedure but as one layer in a larger formal pipeline from debate text to structured reasoning and query answering (Alfano et al., 3 Mar 2026). The supplied material therefore situates QuAD semantics at the intersection of computational argumentation, iterative fixed-point methods, and neuro-symbolic debate analysis.